attention.py 19.7 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
# Copyright 2023 The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14
from typing import Any, Dict, Optional, Tuple
15
16

import torch
Patrick von Platen's avatar
Patrick von Platen committed
17
import torch.nn.functional as F
18
19
from torch import nn

Dhruv Nair's avatar
Dhruv Nair committed
20
from ..utils.torch_utils import maybe_allow_in_graph
21
from .activations import get_activation
Patrick von Platen's avatar
Patrick von Platen committed
22
from .attention_processor import Attention
Kashif Rasul's avatar
Kashif Rasul committed
23
from .embeddings import CombinedTimestepLabelEmbeddings
24
from .lora import LoRACompatibleLinear
25
26


27
28
@maybe_allow_in_graph
class GatedSelfAttentionDense(nn.Module):
29
30
31
32
33
34
35
36
37
38
39
    r"""
    A gated self-attention dense layer that combines visual features and object features.

    Parameters:
        query_dim (`int`): The number of channels in the query.
        context_dim (`int`): The number of channels in the context.
        n_heads (`int`): The number of heads to use for attention.
        d_head (`int`): The number of channels in each head.
    """

    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
        super().__init__()

        # we need a linear projection since we need cat visual feature and obj feature
        self.linear = nn.Linear(context_dim, query_dim)

        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
        self.ff = FeedForward(query_dim, activation_fn="geglu")

        self.norm1 = nn.LayerNorm(query_dim)
        self.norm2 = nn.LayerNorm(query_dim)

        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))

        self.enabled = True

56
    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
57
58
59
60
61
62
63
64
65
66
67
68
        if not self.enabled:
            return x

        n_visual = x.shape[1]
        objs = self.linear(objs)

        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))

        return x


69
@maybe_allow_in_graph
Patrick von Platen's avatar
Patrick von Platen committed
70
class BasicTransformerBlock(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
71
72
73
74
    r"""
    A basic Transformer block.

    Parameters:
Will Berman's avatar
Will Berman committed
75
76
77
78
        dim (`int`): The number of channels in the input and output.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
Will Berman's avatar
Will Berman committed
79
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
Will Berman's avatar
Will Berman committed
80
81
82
83
84
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm (:
            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (:
            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
85
86
87
88
89
90
91
92
93
94
95
96
97
98
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used.
        double_self_attention (`bool`, *optional*):
            Whether to use two self-attention layers. In this case no cross attention layers are used.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
            Whether to use learnable elementwise affine parameters for normalization.
        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
        final_dropout (`bool` *optional*, defaults to False):
            Whether to apply a final dropout after the last feed-forward layer.
        attention_type (`str`, *optional*, defaults to `"default"`):
            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
Kashif Rasul's avatar
Kashif Rasul committed
99
100
101
102
103
    """

    def __init__(
        self,
        dim: int,
Will Berman's avatar
Will Berman committed
104
105
        num_attention_heads: int,
        attention_head_dim: int,
Kashif Rasul's avatar
Kashif Rasul committed
106
        dropout=0.0,
Will Berman's avatar
Will Berman committed
107
108
109
110
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
111
        only_cross_attention: bool = False,
112
        double_self_attention: bool = False,
113
        upcast_attention: bool = False,
Kashif Rasul's avatar
Kashif Rasul committed
114
115
116
        norm_elementwise_affine: bool = True,
        norm_type: str = "layer_norm",
        final_dropout: bool = False,
117
        attention_type: str = "default",
Kashif Rasul's avatar
Kashif Rasul committed
118
    ):
Patrick von Platen's avatar
Patrick von Platen committed
119
        super().__init__()
120
        self.only_cross_attention = only_cross_attention
Kashif Rasul's avatar
Kashif Rasul committed
121
122
123
124
125
126
127
128
129

        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"

        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
            raise ValueError(
                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
            )
130

131
        # Define 3 blocks. Each block has its own normalization layer.
132
        # 1. Self-Attn
133
134
135
136
137
138
        if self.use_ada_layer_norm:
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        elif self.use_ada_layer_norm_zero:
            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
Patrick von Platen's avatar
Patrick von Platen committed
139
        self.attn1 = Attention(
Will Berman's avatar
Will Berman committed
140
141
142
143
144
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
145
            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
146
            upcast_attention=upcast_attention,
147
148
        )

149
        # 2. Cross-Attn
150
        if cross_attention_dim is not None or double_self_attention:
151
152
153
154
155
156
157
158
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
            )
Patrick von Platen's avatar
Patrick von Platen committed
159
            self.attn2 = Attention(
160
                query_dim=dim,
161
                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
162
163
164
165
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
166
                upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
167
            )  # is self-attn if encoder_hidden_states is none
168
169
        else:
            self.norm2 = None
170
            self.attn2 = None
171
172

        # 3. Feed-forward
Kashif Rasul's avatar
Kashif Rasul committed
173
        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
174
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
Patrick von Platen's avatar
Patrick von Platen committed
175

176
        # 4. Fuser
177
        if attention_type == "gated" or attention_type == "gated-text-image":
178
179
            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)

180
181
182
183
184
185
186
187
188
        # let chunk size default to None
        self._chunk_size = None
        self._chunk_dim = 0

    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
        # Sets chunk feed-forward
        self._chunk_size = chunk_size
        self._chunk_dim = dim

189
190
    def forward(
        self,
191
192
193
194
195
196
197
        hidden_states: torch.FloatTensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        class_labels: Optional[torch.LongTensor] = None,
198
    ) -> torch.FloatTensor:
199
        # Notice that normalization is always applied before the real computation in the following blocks.
200
        # 0. Self-Attention
Kashif Rasul's avatar
Kashif Rasul committed
201
202
203
204
205
206
207
208
209
        if self.use_ada_layer_norm:
            norm_hidden_states = self.norm1(hidden_states, timestep)
        elif self.use_ada_layer_norm_zero:
            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
            )
        else:
            norm_hidden_states = self.norm1(hidden_states)

210
211
212
213
        # 1. Retrieve lora scale.
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

        # 2. Prepare GLIGEN inputs
214
215
        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
216

217
218
219
220
221
222
        attn_output = self.attn1(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )
Kashif Rasul's avatar
Kashif Rasul committed
223
224
        if self.use_ada_layer_norm_zero:
            attn_output = gate_msa.unsqueeze(1) * attn_output
225
        hidden_states = attn_output + hidden_states
Will Berman's avatar
Will Berman committed
226

227
        # 2.5 GLIGEN Control
228
229
        if gligen_kwargs is not None:
            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
230
        # 2.5 ends
231

232
        # 3. Cross-Attention
233
234
235
236
        if self.attn2 is not None:
            norm_hidden_states = (
                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
            )
Kashif Rasul's avatar
Kashif Rasul committed
237

238
239
240
            attn_output = self.attn2(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states,
241
                attention_mask=encoder_attention_mask,
242
                **cross_attention_kwargs,
Will Berman's avatar
Will Berman committed
243
            )
244
            hidden_states = attn_output + hidden_states
Will Berman's avatar
Will Berman committed
245

246
        # 4. Feed-forward
Kashif Rasul's avatar
Kashif Rasul committed
247
248
249
250
251
        norm_hidden_states = self.norm3(hidden_states)

        if self.use_ada_layer_norm_zero:
            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]

252
253
254
255
256
257
258
259
260
        if self._chunk_size is not None:
            # "feed_forward_chunk_size" can be used to save memory
            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
                raise ValueError(
                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
                )

            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
            ff_output = torch.cat(
261
262
263
264
                [
                    self.ff(hid_slice, scale=lora_scale)
                    for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
                ],
265
266
267
                dim=self._chunk_dim,
            )
        else:
268
            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
Kashif Rasul's avatar
Kashif Rasul committed
269
270
271
272
273

        if self.use_ada_layer_norm_zero:
            ff_output = gate_mlp.unsqueeze(1) * ff_output

        hidden_states = ff_output + hidden_states
Will Berman's avatar
Will Berman committed
274

275
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
276
277
278


class FeedForward(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
279
280
281
282
    r"""
    A feed-forward layer.

    Parameters:
Will Berman's avatar
Will Berman committed
283
284
285
286
287
        dim (`int`): The number of channels in the input.
        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
Kashif Rasul's avatar
Kashif Rasul committed
288
        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
Kashif Rasul's avatar
Kashif Rasul committed
289
290
291
    """

    def __init__(
Will Berman's avatar
Will Berman committed
292
293
294
295
296
297
        self,
        dim: int,
        dim_out: Optional[int] = None,
        mult: int = 4,
        dropout: float = 0.0,
        activation_fn: str = "geglu",
Kashif Rasul's avatar
Kashif Rasul committed
298
        final_dropout: bool = False,
Kashif Rasul's avatar
Kashif Rasul committed
299
    ):
Patrick von Platen's avatar
Patrick von Platen committed
300
301
        super().__init__()
        inner_dim = int(dim * mult)
302
        dim_out = dim_out if dim_out is not None else dim
Patrick von Platen's avatar
Patrick von Platen committed
303

304
305
        if activation_fn == "gelu":
            act_fn = GELU(dim, inner_dim)
Kashif Rasul's avatar
Kashif Rasul committed
306
307
        if activation_fn == "gelu-approximate":
            act_fn = GELU(dim, inner_dim, approximate="tanh")
308
309
        elif activation_fn == "geglu":
            act_fn = GEGLU(dim, inner_dim)
Will Berman's avatar
Will Berman committed
310
        elif activation_fn == "geglu-approximate":
311
            act_fn = ApproximateGELU(dim, inner_dim)
Will Berman's avatar
Will Berman committed
312
313

        self.net = nn.ModuleList([])
314
        # project in
315
        self.net.append(act_fn)
316
317
318
        # project dropout
        self.net.append(nn.Dropout(dropout))
        # project out
319
        self.net.append(LoRACompatibleLinear(inner_dim, dim_out))
Kashif Rasul's avatar
Kashif Rasul committed
320
321
322
        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
        if final_dropout:
            self.net.append(nn.Dropout(dropout))
Patrick von Platen's avatar
Patrick von Platen committed
323

324
    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
325
        for module in self.net:
326
327
328
329
            if isinstance(module, (LoRACompatibleLinear, GEGLU)):
                hidden_states = module(hidden_states, scale)
            else:
                hidden_states = module(hidden_states)
330
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
331

Patrick von Platen's avatar
Patrick von Platen committed
332

333
334
class GELU(nn.Module):
    r"""
Kashif Rasul's avatar
Kashif Rasul committed
335
    GELU activation function with tanh approximation support with `approximate="tanh"`.
336
337
338
339
340

    Parameters:
        dim_in (`int`): The number of channels in the input.
        dim_out (`int`): The number of channels in the output.
        approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
341
342
    """

Kashif Rasul's avatar
Kashif Rasul committed
343
    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none"):
344
345
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out)
Kashif Rasul's avatar
Kashif Rasul committed
346
        self.approximate = approximate
347

348
    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
349
        if gate.device.type != "mps":
Kashif Rasul's avatar
Kashif Rasul committed
350
            return F.gelu(gate, approximate=self.approximate)
351
        # mps: gelu is not implemented for float16
Kashif Rasul's avatar
Kashif Rasul committed
352
        return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
353
354
355
356
357
358
359

    def forward(self, hidden_states):
        hidden_states = self.proj(hidden_states)
        hidden_states = self.gelu(hidden_states)
        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
360
class GEGLU(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
361
362
363
364
    r"""
    A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.

    Parameters:
Will Berman's avatar
Will Berman committed
365
366
        dim_in (`int`): The number of channels in the input.
        dim_out (`int`): The number of channels in the output.
Kashif Rasul's avatar
Kashif Rasul committed
367
368
369
    """

    def __init__(self, dim_in: int, dim_out: int):
Patrick von Platen's avatar
Patrick von Platen committed
370
        super().__init__()
371
        self.proj = LoRACompatibleLinear(dim_in, dim_out * 2)
Patrick von Platen's avatar
Patrick von Platen committed
372

373
    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
374
375
376
377
378
        if gate.device.type != "mps":
            return F.gelu(gate)
        # mps: gelu is not implemented for float16
        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)

379
380
    def forward(self, hidden_states, scale: float = 1.0):
        hidden_states, gate = self.proj(hidden_states, scale).chunk(2, dim=-1)
381
        return hidden_states * self.gelu(gate)
Will Berman's avatar
Will Berman committed
382
383
384


class ApproximateGELU(nn.Module):
385
386
387
    r"""
    The approximate form of Gaussian Error Linear Unit (GELU). For more details, see section 2:
    https://arxiv.org/abs/1606.08415.
Will Berman's avatar
Will Berman committed
388

389
390
391
    Parameters:
        dim_in (`int`): The number of channels in the input.
        dim_out (`int`): The number of channels in the output.
Will Berman's avatar
Will Berman committed
392
393
394
395
396
397
    """

    def __init__(self, dim_in: int, dim_out: int):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out)

398
    def forward(self, x: torch.Tensor) -> torch.Tensor:
Will Berman's avatar
Will Berman committed
399
400
401
402
403
        x = self.proj(x)
        return x * torch.sigmoid(1.702 * x)


class AdaLayerNorm(nn.Module):
404
    r"""
Will Berman's avatar
Will Berman committed
405
    Norm layer modified to incorporate timestep embeddings.
406
407
408
409

    Parameters:
        embedding_dim (`int`): The size of each embedding vector.
        num_embeddings (`int`): The size of the dictionary of embeddings.
Will Berman's avatar
Will Berman committed
410
411
    """

412
    def __init__(self, embedding_dim: int, num_embeddings: int):
Will Berman's avatar
Will Berman committed
413
414
415
416
417
418
        super().__init__()
        self.emb = nn.Embedding(num_embeddings, embedding_dim)
        self.silu = nn.SiLU()
        self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)

419
    def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:
Will Berman's avatar
Will Berman committed
420
421
422
423
        emb = self.linear(self.silu(self.emb(timestep)))
        scale, shift = torch.chunk(emb, 2)
        x = self.norm(x) * (1 + scale) + shift
        return x
Kashif Rasul's avatar
Kashif Rasul committed
424
425
426


class AdaLayerNormZero(nn.Module):
427
    r"""
Kashif Rasul's avatar
Kashif Rasul committed
428
    Norm layer adaptive layer norm zero (adaLN-Zero).
429
430
431
432

    Parameters:
        embedding_dim (`int`): The size of each embedding vector.
        num_embeddings (`int`): The size of the dictionary of embeddings.
Kashif Rasul's avatar
Kashif Rasul committed
433
434
    """

435
    def __init__(self, embedding_dim: int, num_embeddings: int):
Kashif Rasul's avatar
Kashif Rasul committed
436
437
438
439
440
441
442
443
        super().__init__()

        self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim)

        self.silu = nn.SiLU()
        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)

444
445
446
447
448
449
450
    def forward(
        self,
        x: torch.Tensor,
        timestep: torch.Tensor,
        class_labels: torch.LongTensor,
        hidden_dtype: Optional[torch.dtype] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
Kashif Rasul's avatar
Kashif Rasul committed
451
452
453
454
        emb = self.linear(self.silu(self.emb(timestep, class_labels, hidden_dtype=hidden_dtype)))
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)
        x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]
        return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
455
456
457


class AdaGroupNorm(nn.Module):
458
    r"""
459
    GroupNorm layer modified to incorporate timestep embeddings.
460
461
462
463
464
465
466

    Parameters:
        embedding_dim (`int`): The size of each embedding vector.
        num_embeddings (`int`): The size of the dictionary of embeddings.
        num_groups (`int`): The number of groups to separate the channels into.
        act_fn (`str`, *optional*, defaults to `None`): The activation function to use.
        eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability.
467
468
469
470
471
472
473
474
    """

    def __init__(
        self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5
    ):
        super().__init__()
        self.num_groups = num_groups
        self.eps = eps
475
476
477
478
479

        if act_fn is None:
            self.act = None
        else:
            self.act = get_activation(act_fn)
480
481
482

        self.linear = nn.Linear(embedding_dim, out_dim * 2)

483
    def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:
484
485
486
487
488
489
490
491
492
        if self.act:
            emb = self.act(emb)
        emb = self.linear(emb)
        emb = emb[:, :, None, None]
        scale, shift = emb.chunk(2, dim=1)

        x = F.group_norm(x, self.num_groups, eps=self.eps)
        x = x * (1 + scale) + shift
        return x